research: offline translation model evaluation (VRAM vs accuracy) #18

New issue

Open

opened 2026-04-06 13:11:43 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-06 13:11:43 -07:00

Owner

Research questions:

What is DeepL's offline model VRAM requirement?
Can it coexist with YAMNet + wav2vec2 + whisper.cpp on Heimdall's current VRAM headroom?
If not, what's the smallest OSS translation model that covers top 10 languages at acceptable quality for 5-15 word tone labels?

Output: a recommendation (model name, VRAM requirement, quality notes) that unblocks the offline translation implementation issue.

This is the open question from design doc §Open Questions #1.

Before committing to offline translation as a Free-tier feature, evaluate whether DeepL's offline model (or an OSS alternative like NLLB-200 or Opus-MT) can run alongside cf_voice classifiers on Heimdall's VRAM. Research questions: 1. What is DeepL's offline model VRAM requirement? 2. Can it coexist with YAMNet + wav2vec2 + whisper.cpp on Heimdall's current VRAM headroom? 3. If not, what's the smallest OSS translation model that covers top 10 languages at acceptable quality for 5-15 word tone labels? Output: a recommendation (model name, VRAM requirement, quality notes) that unblocks the offline translation implementation issue. This is the open question from design doc §Open Questions #1.